How to Forecast AI Spend When Nobody Owns the Budget

The budget model that doesn't exist yet

Ask a CFO how much the company will spend on AI next quarter, and you'll get one of two answers: a shrug, or a number that's wrong by 40%. Neither is acceptable when AI is on track to become a top-five line item at most enterprises within 18 months.

The problem isn't that finance teams are bad at forecasting. It's that AI spend doesn't behave like anything they've forecast before. Cloud infrastructure costs are variable but directionally predictable — you can look at trailing compute usage and project forward with reasonable confidence. SaaS costs are seat-based and linear. AI spend is neither. It's a mix of per-seat licenses, consumption-based API billing, embedded AI features buried inside existing SaaS contracts, and shadow subscriptions that don't hit procurement at all.

Why traditional models break

Traditional IT budgeting assumes you can categorize spend into fixed (licenses, contracts) and variable (compute, bandwidth) buckets, then forecast each independently. AI breaks this in three specific ways.

First, token-based costs are non-linear. A team running OpenAI's API for document summarization might spend $800/month for six months straight, then jump to $6,000/month because they deployed an agent that chains multiple API calls per request. The underlying usage didn't grow linearly — the architecture changed. You can't forecast this from a trendline.

Second, seat costs compound silently. A Copilot license looks like a predictable $30–$40/seat/month. But seats grow with headcount, and in many organizations, AI tool provisioning is happening at the team level without centralized approval. Your headcount plan says 200 new hires this year; your AI seat count might grow by 500 because existing employees are adding tools too.

Third, pricing models shift underneath you. AI vendors are still figuring out their economics. OpenAI has changed pricing four times in 18 months. Anthropic's model tiers create step-function cost changes when teams upgrade. Understanding AI pricing models is a prerequisite for forecasting, but even that knowledge has a short shelf life.

A practical forecasting framework

Given these dynamics, here's a framework that actually works. It won't give you the false precision of a traditional budget model, but it will give you a defensible range that holds up in board conversations.

Step 1: Build a complete inventory. You cannot forecast what you cannot see. Catalog every AI tool, API account, and embedded AI feature across the organization. Include the ones on corporate cards. Include the ones buried in existing SaaS contracts. The real number is always higher than IT reports.

Step 2: Classify by cost behavior. Group each tool into one of three buckets: fixed-seat (predictable per user per month), consumption-based (variable by usage), or embedded (included in another contract, cost allocated elsewhere). Each bucket gets a different forecasting approach.

Step 3: Forecast seats with a growth multiplier.Don't just use headcount projections. Apply a 1.5–2.5x multiplier to account for organic tool adoption within existing headcount. Look at trailing three-month seat growth rates by department — some teams (engineering, marketing) adopt faster than others.

Step 4: Forecast consumption with scenario bands. For API-based spend, build three scenarios: baseline (current trajectory), moderate (one new use case per team per quarter), and aggressive (agent deployment or architecture changes). Weight them 40/40/20. The aggressive scenario is where your budget risk lives.

Step 5: Add a shadow AI buffer.Add 15–25% to your total forecast for tools you haven't discovered yet. This isn't padding — it's based on what we consistently see when running full AI spend audits. The gap between known and actual spend is real and persistent.

Who should own this?

The biggest forecasting failure isn't methodological — it's organizational. AI spend typically has no single owner. IT owns some tools. Individual departments own others. Procurement may not even be involved. When nobody owns the budget, nobody forecasts it.

The companies getting this right are treating AI spend the way they eventually learned to treat cloud spend: as a cross-functional discipline that needs a dedicated owner, shared data, and executive sponsorship. Whether that lives in FinOps, FP&A, or a new AI operations function matters less than the fact that someone is accountable for the number.

As we've argued before, AI costs are following the same trajectory as early cloud costs— but compressed into a fraction of the time. The organizations that build forecasting muscle now will avoid the years of overspending that defined the first decade of cloud.

The cost of getting it wrong

A bad AI forecast doesn't just mean a missed budget target. It means the CFO loses confidence in AI investments broadly, which leads to blanket freezes rather than strategic allocation. It means departments hoard their own AI budgets defensively rather than consolidating for better pricing. It means the board gets surprised, and surprised boards don't fund expansion.

The goal isn't perfect accuracy. It's a defensible range that lets the organization invest in AI with eyes open — knowing what the realistic cost trajectory looks like, where the risk concentrates, and what levers exist to control it. Finance teams that build this capability now will be the ones who actually get to say yes to AI initiatives, because they can show the board they understand the cost.